Performance evaluation of saliency map methods on remotely sensed RGB images

Sönmez, Selen
Predictive applications of human eye visualization so called saliency map computational models become more attractive in image processing studies. Saliency map highlights regions that are distinctive from their surrounding in the images in interest. In this study, various computational models for salient region detection are investigated on remotely sensed images. The computational methods considered are Itti-Koch, Graph-Based Visual Saliency, Saliency Detection by Combining Simple Priors, Frequency-tuned Salient Region Detection, Image Signature and Region Covariance based Saliency. For evaluation of the computational methods, a dataset containing 226 remotely sensed RGB images has been prepared. The dataset forestry and water surface images captured in three different levels. The saliency maps produced by the computational methods on the dataset are compared with the saliency maps extracted from data collected in experiment conducted on human subjects. In these experiments 20 subjects are participated and the data is collected by using Tobii T120 Eye Tracker device while the images in the dataset are presented to subjects on computer screen. In the performance evaluation, the saliency maps obtained from human subjects are used as ground truth. The performances of the computational methods are determined by computing similarity of their results to ground truth. As similarity measure, Cosine correlation, Pearson correlation and Structural Similarity index are used. Our experimental evaluation demonstrated that Region Covariance based Saliency and Graph-Based Visual Saliency are the best saliency methods among those that we considered for saliency map generation of remotely sensed RGB images.
Citation Formats
S. Sönmez, “Performance evaluation of saliency map methods on remotely sensed RGB images,” M.S. - Master of Science, Middle East Technical University, 2016.